When Does the US Become a Net Importer of Tokens?

Open original

Tom Tunguz argues that AI demand will outrun domestic data center capacity, making the US a net importer of tokens around 2028 to 2030, depending on how many planned facilities get built. He also says rising compute prices won't cool demand much, so the real limits will be power, capital, and policy.

  • Demand is set to outgrow supply: US data center power goes from 41 GW in 2026 to 66 GW in 2027, while only about 55% of scheduled capacity lands on time.
  • Efficiency gains of roughly 2.9x per year can't fully offset demand growth above about 2.5x annual token growth, so the gap opens by 2028 to 2030.
  • Demand has been price-inelastic: GPU rental prices jumped from $3-5 to $12-13 per hour and demand barely dropped, which means scarcity pricing won't ration it. Capital costs, not electricity, are the bigger constraint on building capacity.

The core claim is that US token demand will outrun domestic capacity sometime between 2028 and 2030, at which point the country becomes a net importer of tokens, much as it already imports rare earths. Tunguz's rough model says current systems plus planned data centers can absorb about 2.5x annual token growth. Faster growth than that, and efficiency gains in chips and models can't keep up with power requirements, since gigawatts needed equals tokens demanded divided by tokens per watt. If 45% of planned projects never connect to the grid, imports start in 2028. Cut cancellations to 20% and the date slips to 2029. Even if every proposed facility gets built on schedule, the gap still opens around 2030. The more interesting part is pricing. Most physical models hold compute prices fixed, but Tunguz points out that demand has so far been stubbornly price-insensitive. Short-term GPU rental prices jumped from roughly $3–$5 an hour to $12–$13 within six months, and buyers, including Anthropic in its burst compute deals, mostly just paid up. He asks where the breaking point is: frontier pretraining budgets, enterprise vibe-coding teams burning hundreds of millions of tokens a week, or consumers paying $40 a month for agent subscriptions. His argument is that scarcity pricing can throttle demand in the short run, but it doesn't change the underlying fact that cheaper production would unlock far more demand, and if the US doesn't build the power and facilities to meet it, the tokens will be produced abroad and imported. The supporting figures are where he gets specific. Goldman Sachs projects US data center power rising from 41 GW in 2026 to 66 GW in 2027, and efficiency is improving about 2.9x per year in tokens per watt, helped by Nvidia's Vera Rubin, Google's Ironwood, and AWS's Trainium3. Only about 55% of scheduled US capacity arrives on time. On cost, power is just 5% of a GPU-hour while capital recovery is 64%, so financing terms matter more than electricity prices. Behind-the-meter gas generation can cover part of the shortfall, but gas turbines are reportedly sold out into 2030, which makes them the likely binding constraint. Tunguz's closing argument is that the limits on data center size, power draw, and location will be set by policy and national priorities rather than by data, algorithms, or GPU supply.

Ask about this article